Development and internal validation of a machine learning-based model to predict 1-year all-cause mortality in patients with gastrointestinal bleeding with type 2 diabetes.

backgroundGastrointestinal bleeding (GIB) is a life-threatening clinical event. Patients with diabetes have a higher risk of GIB. Therefore, it is crucial to predict the 1-year all-cause mortality in patients with type 2 diabetes with GIB.

methodsData from the Medical Information Marketplace for Intensive Care IV, with 1,048 patients were enrolled after applying exclusion criteria. The wrapper method was used to select the optimal subset of features by 10 machine learning learners. These 10 feature subsets in combination with 10 learners were used for model development. Model prediction performance was evaluated using time-dependent concordance index, receiver operating characteristic curve area under the curve (AUC), calibration curves, brier scores, and decision curve analysis in both the training set and the internal validation set. The model was interpreted using the SHapley Additive exPlanations (SHAP) algorithm.

resultsThe Accelerated Oblique Random Survival Forest (AORSF) model had the best predictive performance. The 8 predictors were age, sex, Blood Urea Nitrogen, Red Cell Distribution Width, White Blood Cell, Activated Partial Thromboplastin, Sodium, and Platelet. In the training set, the AUCs at days 60, 180 and 270 were 0.83 (95% CI: 0.79, 0.87), 0.81 (95% CI: 0.77, 0.84), and 0.82 (95% CI: 0.79, 0.86), respectively. The model demonstrated strong calibration, as evidenced by the calibration curves and low Brier scores.

conclusionInterpretable ML models for mortality prediction in diabetic GIB patients are feasible and demonstrate promising predictive performance, which can help clinicians assess disease severity and guide clinical management. CLINICAL TRIAL NUMBER: Not applicable.